Merge pull request #60 from patlux/fix/text-model-image-history

fix(core): omit historical images for text models

CHANGELOG.md

@@ -2,6 +2,7 @@
 
 ## Unreleased
 
+- Allow switching from a vision-capable model to a text-only model by omitting historical image tool results while preserving their text output; direct image prompts still fail clearly.
 - Stream incremental tool-call arguments from the `/alpha/generate` transport instead of waiting for the final complete tool-call event.
 - Add a daily GitHub Actions synchronization job that opens or updates a pull request for CLI version, image capability, reasoning, effort, and output-limit changes in the latest published Command Code catalog.
 - Refresh static model capabilities from `command-code@1.32.2`, separating reasoning support from selectable effort levels and honoring model-specific output limits.

src/converters.ts

@@ -66,9 +66,8 @@ function imageContentError(role: string): Error {
 
 export function assertTextOnlyMessages(messages?: readonly MessageLike[]): void {
   for (const message of messages ?? []) {
-    if (imageParts(message.content).length > 0) {
-      const role = message.role === "toolResult" ? "tool results" : `${message.role} messages`
-      throw imageContentError(role)
+    if (message.role !== "toolResult" && imageParts(message.content).length > 0) {
+      throw imageContentError(`${message.role} messages`)
     }
   }
 }
@@ -268,6 +267,11 @@ export function messagesToCC(
       if (missingResults.length > 0) out.push({ role: "tool", content: missingResults })
     } else if (message.role === "toolResult") {
       if (!message.toolCallId || !callIds.has(message.toolCallId)) continue
+      const images = imageParts(message.content)
+      const text = textContent(message)
+      const outputText =
+        text ||
+        (images.length > 0 && !allowImages ? "[Image omitted: model does not support images]" : "")
       out.push({
         role: "tool",
         content: [
@@ -276,15 +280,13 @@ export function messagesToCC(
             toolCallId: message.toolCallId,
             toolName: message.toolName,
             output: message.isError
-              ? { type: "error-text", value: textContent(message) }
-              : { type: "text", value: textContent(message) },
+              ? { type: "error-text", value: outputText }
+              : { type: "text", value: outputText },
           },
         ],
       })
 
-      const images = imageParts(message.content)
-      if (images.length > 0) {
-        if (!allowImages) throw imageContentError("tool results")
+      if (images.length > 0 && allowImages) {
         out.push({
           role: "user",
           content: images.map(imageToCommandCode),

tests/test-pure-functions.ts

@@ -161,7 +161,7 @@ describe("projectSlugFromPath()", () => {
 })
 
 describe("text-only image handling", () => {
-  it("rejects image content for models without image support", () => {
+  it("rejects direct image input for models without image support", () => {
     assert.throws(
       () =>
         assertTextOnlyMessages([
@@ -172,8 +172,10 @@ describe("text-only image handling", () => {
         ]),
       /does not support image content/i,
     )
-    assert.throws(
-      () =>
+  })
+
+  it("allows historical tool-result images to be omitted for text-only models", () => {
+    assert.doesNotThrow(() =>
       assertTextOnlyMessages([
         {
           role: "toolResult",
@@ -181,7 +183,6 @@ describe("text-only image handling", () => {
           content: [{ type: "image", data: "base64-data", mimeType: "image/png" }],
         },
       ]),
-      /does not support image content/i,
     )
   })
 })
@@ -597,6 +598,48 @@ describe("messagesToCC()", () => {
     )
   })
 
+  it("omits tool-result images for text-only models while preserving their text", () => {
+    const result = messagesToCC([
+      { role: "user", content: "read image" },
+      {
+        role: "assistant",
+        content: [{ type: "toolCall", id: "c1", name: "read", arguments: {} }],
+      },
+      {
+        role: "toolResult",
+        toolCallId: "c1",
+        toolName: "read",
+        content: [
+          { type: "text", text: "image attached" },
+          { type: "image", data: "aGVsbG8=", mimeType: "image/jpeg" },
+        ],
+      },
+    ])
+
+    assert.equal(objectAt(result, ["2", "content", "0", "output", "value"]), "image attached")
+    assert.equal(objectAt(result, ["3"]), undefined)
+  })
+
+  it("describes an omitted image-only tool result for text-only models", () => {
+    const result = messagesToCC([
+      {
+        role: "assistant",
+        content: [{ type: "toolCall", id: "c1", name: "read", arguments: {} }],
+      },
+      {
+        role: "toolResult",
+        toolCallId: "c1",
+        toolName: "read",
+        content: [{ type: "image", data: "aGVsbG8=", mimeType: "image/jpeg" }],
+      },
+    ])
+
+    assert.equal(
+      objectAt(result, ["1", "content", "0", "output", "value"]),
+      "[Image omitted: model does not support images]",
+    )
+  })
+
   it("preserves tool-result images as a following user image message", () => {
     const result = messagesToCC(
       [

tests/test-stream.ts

@@ -240,12 +240,16 @@ describe("streamCommandCode — successful streams", () => {
     )
   })
 
-  it("rejects a tool-result image before network access for text-only models", async () => {
+  it("omits a historical tool-result image after switching to a text-only model", async () => {
+    server.mockResponse({
+      type: "success",
+      events: [JSON.stringify({ type: "finish", finishReason: "stop" })],
+    })
     const { streamCommandCode } = createTestDeps({ apiBase: server.baseUrl() })
 
     const events = await collectEvents(
       streamCommandCode(
-        makeModel({ id: "deepseek/deepseek-v4-pro" }),
+        makeModel({ id: "deepseek/deepseek-v4-flash" }),
         makeContext({
           messages: [
             { role: "user", content: "read the image" },
@@ -257,19 +261,29 @@ describe("streamCommandCode — successful streams", () => {
               role: "toolResult",
               toolCallId: "c1",
               toolName: "read",
-              content: [{ type: "image", data: "aGVsbG8=", mimeType: "image/png" }],
+              content: [
+                { type: "text", text: "image attached" },
+                { type: "image", data: "aGVsbG8=", mimeType: "image/png" },
+              ],
             },
+            { role: "user", content: "continue without the image" },
           ],
         }),
         { apiKey: "mock-key" },
       ),
     )
 
-    const lastEvent = events.at(-1)
-    assert.equal(lastEvent?.type, "error")
-    if (lastEvent?.type !== "error") throw new Error("expected error event")
-    assert.match(lastEvent.error.errorMessage ?? "", /does not support image content/i)
-    assert.equal(server.requestCount(), 0)
+    assert.equal(events.at(-1)?.type, "done")
+    assert.equal(server.requestCount(), 1)
+    const body = server.lastRequestBody()
+    assert.equal(
+      objectAt(body, ["params", "messages", "2", "content", "0", "output", "value"]),
+      "image attached",
+    )
+    assert.equal(
+      objectAt(body, ["params", "messages", "3", "content"]),
+      "continue without the image",
+    )
   })
 
   it("rejects images before network access for text-only models", async () => {